■This model is experimental and very difficult!

Don't get too depressed if you don't get good results!

Please download these two models.

Pruned Model bf16 (1.92 GB):U-net

Pruned Model fp16 (1.41 GB): Text encoder

Please download the comfy ui workflow if you also need it.

Training Data (2.19 KB):comfy ui workflow

■I replaced it with the text encoder of Animagine-XL 3.1 and redid the fine tuning of u-net.

I've become more acquainted with anime styles than before, and I think the recognition rate for tags has also improved.

I don't know if the quality is better than before...

But it's an interesting experiment, so I'd like to share it with you.

■If you do not want a realistic image to appear, it will be improved by adding "blush" in the positive and "realistic" in the negative.

However, from my experience, flat patterns often break down and the results are not very good.Cascade is not good at handling fingers and the human body, so it does not work as well as SDXL.

Personally, I think it's better to generate 2.5D or realistic images to get the beautiful contrast and background of the cascade.

I prefer to keep negative prompts to a minimum. I don't want to limit the possibilities of the model...

■I haven't learned the quality tags or character tags, but I understand the text encoder, so it may be effective.

Maybe there is a tag rule effect of anime 3.1...I try to put the quality tag at the beginning. ↓

Positive prompt:

masterpiece, best quality, very aesthetic, absurdes,

Negative prompt:

low quality, worst quality, normal quality,lowres,pattern,simple background,realistic,

\__________________________________________________________

Attention This model is very difficult!

I trained using onetrainer. I'm training on a 768px dataset.

Fine-tuning is performed on a 60,000-image dataset that mainly contains anime images, but also some realistic and AI images.

There are cases where the look of something realistic or AI comes out strongly.

It might be a good idea to add "realistic" to the negative prompt.

"blush" This tag may be effective as it forces an anime style.

This is a very strong tag, so putting it near the beginning may be too strong.

If you've created a prompt that you're happy with, but you want more variety and the look is too flat, remove it.

My personal opinion is that negative prompts should be kept to a minimum for a high-quality base model. It would be a waste to limit the possibilities.

On the other hand, it might be fun to try something other than anime.

New discoveries are made in areas that were not originally intended.

I think cascade_stage_c_lite falls into the high quality category as a base model and can generate most basic things.I can roughly understand that common tags are pre-learned.

It's truly beautiful when done well!

It would be interesting to generate various tags using something that can automatically generate tags.

■There is no consistency in style.The quality is poor and there are no fixed settings or prompts.

It has no advantage over existing models and has a narrower dataset.

The advantage is that it is lightweight.

Even though the resolution is high, details such as hands and other small parts are not generated well, which is a characteristic of the model and difficult to improve. It would be fun to consider other approaches such as cascading the composition and enhancing the details with i2i.

■I am training with the danbooru tag.

We are only learning general tags such as 1gril, and we are not training artist or anime work tags.

A small number of tags will produce a disastrous result.The tags that are often used in danbooru and SD are the quality tags for this model.

The order of tags is important. Every tag has a unique image.

The more popular the tag, the better the quality may be, but the image will be reflected more strongly, so it is also effective to offset it with other tags or change the order to dilute it.

If the effect is too strong, it might be a good idea to lower the weight.

"Looking at viewer","upper body","shiny skin"etc... can easily be of high quality.

I'm training without adding the "nsfw" tag, but I feel like it's effective for some reason...

■I've also added a simple comfyui workflow that I'm using for my generation.

I am generating images between 1024-2048px.

Many people have shared their workflows. I'm still in the exploration stage. I think I'll make new discoveries by trying out various settings.

I find it difficult to generate details with cascade. It would be interesting to generate a rough composition with cascade and do i2i with other SD models!The advantage of stage_c_lite is that it is lightweight, so the workflow up to i2i will be less stressful. Me too. I'm enjoying the changes with i2i on my playground-v2-512px!

Generate 1024x1536px with cascade, downscale by 0.75, then i2i at 768x1152 to generate the best of both worlds without the stress of speed.

Ideal composition atmosphere + improved details such as eyes

I would be happy if you could share the results of combinations with other models such as i2i!

You can use it by simply changing stage_c of the existing workflow to this.

I use "madebyollin/stage-a-ft-hq" for stageA when generating. It may reduce noise. I like this model.

■It's an incomplete and very difficult model, but if you're interested, please give it a try.